1906.4010.md

Flair: an easy-to-use framework for stat-of-the-art NLP

Paper, Tags: #nlp, #frameworks

We present Flair, an NLP framework which is simple, unified interface for conceptually very different types of word and document embeddings. Standard model training and hyperparameter selection routines are also implemented.

Previous work

• Hierarchical architectures: combine classic word embeddings with character-level features trained on task data, the output states are concatenated with the output of the embedding layer.
• Contextualized embeddings: different embeddings for the same word depending on its contextual usage (Peters et al., 2018a, Akbik et al., 2018, Devlin et al., 2018). The architecture must have the output states of a trained LM, concatenated with the output of the embedding layer, thus adding architectural complexity.

Word embeddings cannot simply be mixed and matched with minimal effort, they require specific reworkings of the model architecture.

Flair framework

Goal is to abstract away from specific engineering challenges that different types of word embeddings raise. The user can make use of any type of word embedding with no additional engineering effort. We also include a growing list of pre-trained models.

Framework embeddings

• Classic word embeddings
• Other word embeddings: ELMo, BERT, flair (see references above)
• Stacked embeddings
• Document embeddings

Code in Github